PaddleX
Open-source AI development toolkit built on PaddlePaddle, with pretrained OCR and vision models, local CPU or GPU inference, and self-hosted deployment.
PaddleX is a low-code AI development toolkit for developers building document processing, computer vision and time-series applications on their own hardware. Built on PaddlePaddle, it combines pretrained models with tools for training, inference and deployment. It's open source under Apache 2.0.
Document processing is a substantial part of the toolkit. It supports PaddleOCR-VL and multilingual PP-OCRv5 models, alongside pipelines for reading text, recognizing tables, parsing page layouts and extracting information from documents. Formula recognition and seal text recognition cover more specialized inputs. Document preprocessing can correct orientation and straighten warped page images, while table recognition can return reconstructed tables as HTML.
For image applications, PaddleX includes classification, object detection and segmentation, with specialized capabilities such as small-object detection, image anomaly detection and human keypoint detection. Its time-series tools cover forecasting and anomaly detection. Developers can use complete pipelines or combine individual model modules, and the development tools include both Python interfaces and a graphical interface.
Models can run locally on CPUs or NVIDIA GPUs, with support for Kunlunxin, Ascend and Cambricon hardware as well. Linux and Windows are supported. Deployment choices include self-hosted services through Docker and execution on edge devices. Paddle Inference and ONNX Runtime are available as inference backends, and pipeline benchmarks measure total processing time as well as time spent in individual modules.
